scBench
Science · 2026-02-06
scBench contains 195 verifiable agent tasks drawn from practical single-cell RNA-seq workflows. Deterministic graders report the percentage of tasks passed.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 60.6 |
| Claude Mythos 5 | 59.3 |
| Opus 4.8 | 58.2 |
| GPT-5.5 | 58.0 |
| GPT-5.4 | 57.4 |
| Gemini 3.5 Flash | 56.9 |
| Sonnet 5 | 56.2 |
| Opus 4.7 | 55.2 |
| Gemini 3.1 Pro Preview | 53.9 |
| Opus 4.6 | 52.6 |
| GPT-5.2 | 52.3 |
| Sonnet 4.6 | 50.3 |
| Opus 4.5 | 47.2 |
| Grok 4.20 | 44.4 |
| Grok 4.3 | 44.3 |
| GPT-5.1 | 38.8 |
| Sonnet 4.5 | 33.2 |
| Grok 4.1 Fast | 30.3 |
| Gemini 2.5 Pro | 23.6 |